LSTM-Based Lower Limbs Motion Reconstruction Using Low-Dimensional Input of Inertial Motion Capture System

Motion capture system has been widely used in virtual reality and rehabilitation area. This study proposed a data-driven method using low-dimensional input of inertial motion capture system to reconstruct human lower-limb motions. The long short-term memory (LSTM) neural network was used and an ensemble LSTM architecture was involved to improve reconstruction performance. Besides, the selection of optimal sensor configuration scheme and time-step parameters of LSTM network was discussed in detail. The reconstruction experiment shows that the method could get the lowest reconstruction joint angle root mean square (RMS) errors of 4.031° on separated motion dataset, and 5.105° on completely new dataset of synthetic motions using ensemble LSTM model with 18 base learner and three sensors units. The computational consumption test shows that the single and ensemble LSTM model spend 0.15ms and 0.91ms respectively to predict next frame. These findings demonstrate that the proposed method is effective and efficient for motions reconstruction of lower limbs.